Skip to content

Token Preservation Backfire

A token preservation instruction creates a competing objective the agent resolves by doing less work, not by completing the task better.

The Pattern

Add instructions like "preserve tokens," "avoid waste," or "be efficient" to system prompts. The intent is cost savings. The effect is reduced output quality.

Why It Fails

Efficiency instructions create a second objective: minimise resource use. When this competes with the user's task objective, the agent resolves the conflict by doing less work -- refusing explorations, skipping file reads, and stopping early.

Cursor discovered this during their Codex model harness development. GPT-5-Codex, instructed to "preserve tokens and not be wasteful," would sometimes stop with:

"I'm not supposed to waste tokens, and I don't think it's worth continuing with this task!"

The model treated token conservation as a goal in its own right. Rather than optimising how it worked, the instruction changed whether it worked on substantial problems.

flowchart LR
    A["System prompt:<br/>'preserve tokens'"] --> B["Agent receives<br/>two objectives"]
    B --> C["1. Complete the task"]
    B --> D["2. Minimise tokens"]
    C ---|conflict| D
    D --> E["Agent resolves conflict<br/>by doing less"]
    E --> F["Skips exploration"]
    E --> G["Refuses ambitious tasks"]
    E --> H["Stops early"]

The Mechanism

System-level instructions override user-level task requests. When token preservation is a system directive, the efficiency constraint takes precedence over the user's objective — the agent is not being lazy, it is faithfully following a conflicting instruction.

Any instruction framing work as a cost to be minimised risks reducing agent ambition — a form of objective drift where the resource budget displaces the task goal. The effect is most documented for long-horizon coding agents; evidence for other task types is limited to a small number of practitioner reports.

When This Applies

The failure mode is specific to long-horizon, tool-using tasks where the agent chooses whether to explore or continue — coding and file-system work.

Brevity framing stays safe for conversational assistants, summarisation, and single-turn tasks without tool use — where the model has no opportunity to do less work.

The backfire is not universal — a bounded budget differs from an open-ended "don't waste tokens" directive. The Token-Budget-Aware LLM Reasoning framework reports a 68% token reduction with under 5% accuracy loss by inserting an estimated budget into the prompt (arxiv 2412.18547, ACL 2025 Findings; code). The failure is a property of vague resource-minimisation framing, not of efficiency goals as such.

Mitigation

Instead of Use
"Preserve tokens" "Be thorough"
"Don't waste resources" "Bias to action"
"Be efficient and concise" "Implement with reasonable assumptions"
"Minimise tool calls" "Use the tools needed to verify your work"
"Only read files when necessary" "Read files to build context before acting"

Reframe constraints as quality targets rather than resource limits.

Frame around action. OpenAI's Codex prompting guide: "Bias to action: default to implementing with reasonable assumptions; do not end on clarifications unless truly blocked."

Use completion criteria. LangChain structures agent phases (Planning, Build, Verify, Fix) with pre-completion checklists — done means quality criteria met, not budget hit.

Make constraints mechanical. Anthropic recommends requiring absolute filepaths rather than instructing "don't use relative paths" — the constraint enforces itself.

Sources

Key Takeaways

  • Open-ended efficiency instructions ("preserve tokens", "don't be wasteful") create a second objective that long-horizon agents resolve by doing less work — skipping exploration, refusing ambitious tasks, stopping early.
  • The mechanism is instruction precedence: a system-level resource constraint outranks the user's task, so the agent is faithfully following a conflicting directive, not being lazy.
  • The failure mode is specific to multi-step, tool-using tasks where the agent chooses whether to continue (harness-engineering territory); single-turn and summarisation work has no "less work" to fall back to.
  • A bounded, quantified token budget (e.g. TALE) can cut tokens with minimal accuracy loss — the backfire is a property of vague minimisation framing, not of efficiency goals.
  • Reframe constraints as quality targets ("be thorough", "bias to action") or make them mechanical, rather than as resource limits.
Established — multiple independent sources Last reviewed: 2026-06-13
Feedback